Research map / Detection and segmentation / Open-set and few-shot detection
SEEN-DA: SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SEEN-DA, a VLM-based domain adaptive object detector that uses a semantic-entropy-guided domain-aware attention module (inter- and intra-domain branches) inserted into a frozen RegionCLIP visual encoder to refine visual features.
Problem
Domain adaptive object detection: existing methods either ignore semantic information when learning visual features or freeze the visual encoder, so they retain redundant information and discard domain-specific features that matter for target-domain detection.
Approach
Define semantic entropy as the entropy of the softmax distribution between visual features and domain-aware CLIP text embeddings, and use the negated entropy as attention. An inter-domain branch (shared text prompts) extracts domain-invariant features, while an intra-domain branch (separate per-domain learnable prompts and SEEN modules) adds domain-specific semantics, attached as lightweight modul
Key results
- Achieves 57.5% mAP on Cityscapes→Foggy Cityscapes, surpassing DA-Pro (55.9%) by 1.6 points.
- Reaches 67.1% mAP on KITTI→Cityscapes and 66.8% on SIM10K→Cityscapes, the best reported in those tables.
- Reaches 47.9% mAP on Pascal VOC→Clipart, versus 46.7% for SIGMA++ in the table.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SEEN-DA instead of DA-Pro | RegionCLIP (ResNet-50) · Cityscapes→Foggy Cityscapes | mAP: 55.9 -> 57.5 | Helped |
| use SEEN-DA instead of DA-Pro | RegionCLIP (ResNet-50) · KITTI→Cityscapes | mAP: 61.4 -> 67.1 | Helped |
| use SEEN-DA instead of SOCCER | RegionCLIP (ResNet-50) · SIM10K→Cityscapes | mAP: 63.8 -> 66.8 | Helped |
| use SEEN-DA instead of CMT | RegionCLIP (ResNet-50) · Pascal VOC→Clipart | mAP: 47.0 -> 47.9 | Helped |
| replace self-attention with domain-aware attention (ablation) | SEEN-DA · Cross-Weather, Cross-FoV, Sim-to-Real, and Cross-Style | mAP: 54.8 -> 57.5; 61.9 -> 67.1; 62.3 -> 66.8; 43.1 -> 47.9 | Helped |
| replace cross-attention with domain-aware attention (ablation) | SEEN-DA · Cross-Weather, Cross-FoV, Sim-to-Real, and Cross-Style | mAP: 55.5 -> 57.5; 63.5 -> 67.1; 63.7 -> 66.8; 45.2 -> 47.9 | Helped |
| add inter-domain attention (ablation) | SEEN-DA · Cross-Weather | mAP: 52.6 -> 54.9 (+2.3) | Helped |
| add adversarial loss Ladv (ablation) | SEEN-DA with inter-domain attention · Cross-Weather | mAP: 54.9 -> 55.8 (+0.9) | Helped |
| add intra-domain attention (ablation) | SEEN-DA with inter-domain attention and Ladv · Cross-Weather | mAP: 55.8 -> 57.5 (+1.7) | Helped |
| project visual embedding into textual space (V2T) rather than use T2V or T&V projection (ablation) | SEEN-DA SEEN module · Cross-Weather | mAP: 55.1; 2.1%; 57.5% | Helped |
Surprising findings
- Projecting visual embeddings into the textual space worked better (57.5%) than projecting textual embeddings into visual space (55.1%).
Limitations
- Visual encoder is frozen except for lightweight attached attention modules, which limits adaptation capacity.
- Target-domain supervision relies on pseudo labels generated with a fixed prompt about foggy weather, so it is tailored to the cross-weather setting.
Experimental setup
Datasets: Cityscapes Foggy Cityscapes KITTI SIM10K Pascal VOC Clipart
Benchmarks: Cityscapes→Foggy Cityscapes KITTI→Cityscapes SIM10K→Cityscapes Pascal VOC→Clipart
Metrics: mAP@0.5
Baselines: DA-Faster SIGMA++ CIGAR OADA SOCCER DSD-DA CAT NSA-UDA REACT DA-Pro RegionCLIP AT
Builds on: DA-Pro RegionCLIP Faster R-CNN CLIP
Closest papers by meaning
- DA-Ada: Learning Domain-Aware Adapter for Domain Adaptive Object Detection (NeurIPS 2024 · similarity 0.93)
- Universal Domain Adaptive Object Detection via Dual Probabilistic Alignment (AAAI 2025 · similarity 0.90)
- Large Self-Supervised Models Bridge the Gap in Domain Adaptive Object Detection (CVPR 2025 · similarity 0.88)
- Non-exemplar Domain Incremental Object Detection via Learning Domain Bias (AAAI 2024 · similarity 0.88)
- Towards Unsupervised Model Selection for Domain Adaptive Object Detection (NeurIPS 2024 · similarity 0.87)
- Triple Feature Disentanglement for One-Stage Adaptive Object Detection (AAAI 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.